Yilun Zhu

Xiaolin Liu, Yilun Zhu, Xiangyu Zhao et al.

Video multimodal large language models (MLLMs) have made rapid progress on general and long-form video understanding, yet their ability to preserve brief answer-critical visual evidence remains underexplored. Many practical questions are determined by momentary visual events: localized actions or state transitions that may last only a few frames. Such evidence can be skipped by sparse frame sampling, suppressed by visual-token compression, or diluted by coarse temporal aggregation, causing failures that language-side reasoning cannot reliably recover. We introduce Moment-Video, a benchmark for diagnosing the temporal fidelity of video MLLMs through momentary visual event understanding. Each question is grounded in a localized, visually observable, and sampling-sensitive event, requiring models to notice, count, describe, or reason about transient evidence rather than rely on persistent objects, global scene context, or language priors. Moment-Video contains 1,000 human-verified video-QA pairs across 7 domains and 25 fine-grained subcategories, covering four task types: Temporal Occurrence, Temporal Counting, Action Description, and Temporal Reasoning. We evaluate 33 proprietary and open-source MLLMs on Moment-Video. The best-performing model, Seed-2.0-Pro, achieves only 39.6% overall accuracy, while most open-source models remain below 25%, revealing a substantial gap in momentary visual event understanding. Diagnostic analyses show that denser frame sampling improves some models but does not eliminate the bottleneck, and longer videos introduce stronger temporal-localization challenges. These findings suggest that current video MLLMs still lack temporally faithful representations for capturing, preserving, and using brief but decisive visual evidence.

Zdeněk Žabokrtský, Miloslav Konopík, Anna Nedoluzhko et al.

This paper presents an overview of the shared task on multilingual coreference resolution associated with the CRAC 2022 workshop. Shared task participants were supposed to develop trainable systems capable of identifying mentions and clustering them according to identity coreference. The public edition of CorefUD 1.0, which contains 13 datasets for 10 languages, was used as the source of training and evaluation data. The CoNLL score used in previous coreference-oriented shared tasks was used as the main evaluation metric. There were 8 coreference prediction systems submitted by 5 participating teams; in addition, there was a competitive Transformer-based baseline system provided by the organizers at the beginning of the shared task. The winner system outperformed the baseline by 12 percentage points (in terms of the CoNLL scores averaged across all datasets for individual languages).

Yilun Zhu, Aaron Fjeldsted, Darren Holland et al.

The task of mixture proportion estimation (MPE) is to estimate the weight of a component distribution in a mixture, given observations from both the component and mixture. Previous work on MPE adopts the irreducibility assumption, which ensures identifiablity of the mixture proportion. In this paper, we propose a more general sufficient condition that accommodates several settings of interest where irreducibility does not hold. We further present a resampling-based meta-algorithm that takes any existing MPE algorithm designed to work under irreducibility and adapts it to work under our more general condition. Our approach empirically exhibits improved estimation performance relative to baseline methods and to a recently proposed regrouping-based algorithm.

Yilun Zhu, Yuan Zhuang, Nikhita Vedula et al.

Many applications of LLM-based text regression require predicting a full conditional distribution rather than a single point value. We study distributional regression under empirical-quantile supervision, where each input is paired with multiple observed quantile outcomes, and the target distribution is represented by a dense grid of quantiles. We address two key limitations of current approaches: the lack of local grounding for distribution estimates, and the reliance on shared representations that create an indirect bottleneck between inputs and quantile outputs. In this paper, we introduce Quantile Token Regression, which, to our knowledge, is the first work to insert dedicated quantile tokens into the input sequence, enabling direct input-output pathways for each quantile through self-attention. We further augment these quantile tokens with retrieval, incorporating semantically similar neighbor instances and their empirical distributions to ground predictions with local evidence from similar instances. We also provide the first theoretical analysis of loss functions for quantile regression, clarifying which distributional objectives each optimizes. Experiments on the Inside Airbnb and StackSample benchmark datasets with LLMs ranging from 1.7B to 14B parameters show that quantile tokens with neighbors consistently outperform baselines (~4 points lower MAPE and 2x narrower prediction intervals), with especially large gains on smaller and more challenging datasets where quantile tokens produce substantially sharper and more accurate distributions.

Tatsuya Aoyama, Shabnam Behzad, Luke Gessler et al.

We present GENTLE, a new mixed-genre English challenge corpus totaling 17K tokens and consisting of 8 unusual text types for out-of domain evaluation: dictionary entries, esports commentaries, legal documents, medical notes, poetry, mathematical proofs, syllabuses, and threat letters. GENTLE is manually annotated for a variety of popular NLP tasks, including syntactic dependency parsing, entity recognition, coreference resolution, and discourse parsing. We evaluate state-of-the-art NLP systems on GENTLE and find severe degradation for at least some genres in their performance on all tasks, which indicates GENTLE's utility as an evaluation dataset for NLP systems.

Yilun Zhu, Siyao Peng, Sameer Pradhan et al.

Previous attempts to incorporate a mention detection step into end-to-end neural coreference resolution for English have been hampered by the lack of singleton mention span data as well as other entity information. This paper presents a coreference model that learns singletons as well as features such as entity type and information status via a multi-task learning-based approach. This approach achieves new state-of-the-art scores on the OntoGUM benchmark (+2.7 points) and increases robustness on multiple out-of-domain datasets (+2.3 points on average), likely due to greater generalizability for mention detection and utilization of more data from singletons when compared to only coreferent mention pair matching.

Yilun Zhu, Joel Ruben Antony Moniz, Shruti Bhargava et al.

Understanding context is key to understanding human language, an ability which Large Language Models (LLMs) have been increasingly seen to demonstrate to an impressive extent. However, though the evaluation of LLMs encompasses various domains within the realm of Natural Language Processing, limited attention has been paid to probing their linguistic capability of understanding contextual features. This paper introduces a context understanding benchmark by adapting existing datasets to suit the evaluation of generative models. This benchmark comprises of four distinct tasks and nine datasets, all featuring prompts designed to assess the models' ability to understand context. First, we evaluate the performance of LLMs under the in-context learning pretraining scenario. Experimental results indicate that pre-trained dense models struggle with understanding more nuanced contextual features when compared to state-of-the-art fine-tuned models. Second, as LLM compression holds growing significance in both research and real-world applications, we assess the context understanding of quantized models under in-context-learning settings. We find that 3-bit post-training quantization leads to varying degrees of performance reduction on our benchmark. We conduct an extensive analysis of these scenarios to substantiate our experimental results.

Yilun Zhu, Nikhita Vedula, Shervin Malmasi

Entity search, i.e., finding the most similar entities to a query entity, faces unique challenges in e-commerce, where product similarity varies across categories and contexts. Traditional embedding-based approaches often struggle to capture nuanced context-specific attribute relevance. In this paper, we present a two-stage approach combining Large Language Model (LLM)-driven attribute graph construction with graph-aware LLM ranking. In the offline stage, we extract structured product attributes from unstructured text, and construct a reusable attribute graph with category-aware schemas. In the online stage, we rank retrieved candidates by reasoning over this structured representation rather than raw text, reducing per-product token usage by 57% while improving ranking precision. Experiments show that our approach outperforms multiple baselines under zero-shot scenarios, achieving a over 5% improvement in average precision without requiring training data, generalizes robustly across diverse product categories, and shows immense potential for real-world deployment.

Theodore Glavas, Nikhita Vedula, Dushyanta Dhyani et al.

Some text generation tasks, such as Attribute Value Extraction (AVE), require decoding multiple independent sequences from the same document context. While standard autoregressive decoding is slow due to its sequential nature, the independence between output sequences offers an opportunity for parallelism. We present Hyper-Parallel Decoding, a novel decoding algorithm that accelerates offline decoding by leveraging both shared memory and computation across batches. HPD enables out-of-order token generation through position ID manipulation, significantly improving efficiency. Experiments on AVE show that attribute-value pairs are conditionally independent, enabling us to parallelize value generation within each prompt. By further stacking multiple documents within a single prompt, we can decode in parallel up to 96 tokens per prompt. HPD works with all LLMs, and reduces both inference costs and total inference time by up to 13.8X without compromising output quality, potentially saving hundreds of thousands of dollars on industry AVE tasks. Although designed for attribute extraction, HPD makes no assumptions unique to the AVE domain and can in theory be applied to other scenarios with independent output structures.

Yilun Zhu, Jianxin Zhang, Aditya Gangrade et al.

We establish a new theoretical framework for learning under multi-class, instance-dependent label noise. This framework casts learning with label noise as a form of domain adaptation, in particular, domain adaptation under posterior drift. We introduce the concept of \emph{relative signal strength} (RSS), a pointwise measure that quantifies the transferability from noisy to clean posterior. Using RSS, we establish nearly matching upper and lower bounds on the excess risk. Our theoretical findings support the simple \emph{Noise Ignorant Empirical Risk Minimization (NI-ERM)} principle, which minimizes empirical risk while ignoring label noise. Finally, we translate this theoretical insight into practice: by using NI-ERM to fit a linear classifier on top of a self-supervised feature extractor, we achieve state-of-the-art performance on the CIFAR-N data challenge.

Yang Janet Liu, Tatsuya Aoyama, Wesley Scivetti et al.

Work on shallow discourse parsing in English has focused on the Wall Street Journal corpus, the only large-scale dataset for the language in the PDTB framework. However, the data is not openly available, is restricted to the news domain, and is by now 35 years old. In this paper, we present and evaluate a new open-access, multi-genre benchmark for PDTB-style shallow discourse parsing, based on the existing UD English GUM corpus, for which discourse relation annotations in other frameworks already exist. In a series of experiments on cross-domain relation classification, we show that while our dataset is compatible with PDTB, substantial out-of-domain degradation is observed, which can be alleviated by joint training on both datasets.

Yilun Zhu, Nikhita Vedula, Shervin Malmasi

Search queries with superlatives (e.g., best, most popular) require comparing candidates across multiple dimensions, demanding linguistic understanding and domain knowledge. We show that LLMs can uncover latent intent behind these expressions in e-commerce queries through a framework that extracts structured interpretations or hints. Our approach decomposes queries into attribute-value hints generated concurrently with retrieval, enabling efficient integration into the ranking pipeline. Our method improves search performanc eby 10.9 points in MAP and ranking by 5.9 points in MRR over baselines. Since direct LLM-based reranking faces prohibitive latency, we develop an efficient approach transferring superlative interpretations to lightweight models. Our findings provide insights into how superlative semantics can be represented and transferred between models, advancing linguistic interpretation in retrieval systems while addressing practical deployment constraints.

Yixuan Florence Wu, Yilun Zhu, Lei Cao et al.

We propose a new method for statistical inference in generalized linear models. In the overparameterized regime, Principal Component Regression (PCR) reduces variance by projecting high-dimensional data to a low-dimensional principal subspace before fitting. However, PCR incurs truncation bias whenever the true regression vector has mass outside the retained principal components (PC). To mitigate the bias, we propose Calibrated Principal Component Regression (CPCR), which first learns a low-variance prior in the PC subspace and then calibrates the model in the original feature space via a centered Tikhonov step. CPCR leverages cross-fitting and controls the truncation bias by softening PCR's hard cutoff. Theoretically, we calculate the out-of-sample risk in the random matrix regime, which shows that CPCR outperforms standard PCR when the regression signal has non-negligible components in low-variance directions. Empirically, CPCR consistently improves prediction across multiple overparameterized problems. The results highlight CPCR's stability and flexibility in modern overparameterized settings.

Yiheng Chen, Lingyao Li, Zihui Ma et al.

Effective disaster response is essential for safeguarding lives and property. Existing statistical approaches often lack semantic context, generalize poorly across events, and offer limited interpretability. While Large language models (LLMs) provide few-shot generalization, they remain text-bound and blind to geography. To bridge this gap, we introduce a Geospatial Awareness Layer (GAL) that grounds LLM agents in structured earth data. Starting from raw wildfire detections, GAL automatically retrieves and integrates infrastructure, demographic, terrain, and weather information from external geodatabases, assembling them into a concise, unit-annotated perception script. This enriched context enables agents to produce evidence-based resource-allocation recommendations (e.g., personnel assignments, budget allocations), further reinforced by historical analogs and daily change signals for incremental updates. We evaluate the framework in real wildfire scenarios across multiple LLM models, showing that geospatially grounded agents can outperform baselines. The proposed framework can generalize to other hazards such as floods and hurricanes.

Yilun Zhu, Naihao Deng, Naichen Shi et al.

Domain generalization (DG) is the problem of generalizing from several distributions (or domains), for which labeled training data are available, to a new test domain for which no labeled data is available. A common finding in the DG literature is that it is difficult to outperform empirical risk minimization (ERM) on the pooled training data. In this work, we argue that this finding has primarily been reported for datasets satisfying a \emph{covariate shift} assumption. When the dataset satisfies a \emph{posterior drift} assumption instead, we show that ``domain-informed ERM,'' wherein feature vectors are augmented with domain-specific information, outperforms pooling ERM. These claims are supported by a theoretical framework and experiments on language and vision tasks.

Yilun Zhu, Siyao Peng, Sameer Pradhan et al.

Singleton mentions, i.e.~entities mentioned only once in a text, are important to how humans understand discourse from a theoretical perspective. However previous attempts to incorporate their detection in end-to-end neural coreference resolution for English have been hampered by the lack of singleton mention spans in the OntoNotes benchmark. This paper addresses this limitation by combining predicted mentions from existing nested NER systems and features derived from OntoNotes syntax trees. With this approach, we create a near approximation of the OntoNotes dataset with all singleton mentions, achieving ~94% recall on a sample of gold singletons. We then propose a two-step neural mention and coreference resolution system, named SPLICE, and compare its performance to the end-to-end approach in two scenarios: the OntoNotes test set and the out-of-domain (OOD) OntoGUM corpus. Results indicate that reconstructed singleton training yields results comparable to end-to-end systems for OntoNotes, while improving OOD stability (+1.1 avg. F1). We conduct error analysis for mention detection and delve into its impact on coreference clustering, revealing that precision improvements deliver more substantial benefits than increases in recall for resolving coreference chains.

Yilun Zhu, Sameer Pradhan, Amir Zeldes

SOTA coreference resolution produces increasingly impressive scores on the OntoNotes benchmark. However lack of comparable data following the same scheme for more genres makes it difficult to evaluate generalizability to open domain data. Zhu et al. (2021) introduced the creation of the OntoGUM corpus for evaluating geralizability of the latest neural LM-based end-to-end systems. This paper covers details of the mapping process which is a set of deterministic rules applied to the rich syntactic and discourse annotations manually annotated in the GUM corpus. Out-of-domain evaluation across 12 genres shows nearly 15-20% degradation for both deterministic and deep learning systems, indicating a lack of generalizability or covert overfitting in existing coreference resolution models.

Luke Gessler, Shabnam Behzad, Yang Janet Liu et al.

This paper describes our submission to the DISRPT2021 Shared Task on Discourse Unit Segmentation, Connective Detection, and Relation Classification. Our system, called DisCoDisCo, is a Transformer-based neural classifier which enhances contextualized word embeddings (CWEs) with hand-crafted features, relying on tokenwise sequence tagging for discourse segmentation and connective detection, and a feature-rich, encoder-less sentence pair classifier for relation classification. Our results for the first two tasks outperform SOTA scores from the previous 2019 shared task, and results on relation classification suggest strong performance on the new 2021 benchmark. Ablation tests show that including features beyond CWEs are helpful for both tasks, and a partial evaluation of multiple pre-trained Transformer-based language models indicates that models pre-trained on the Next Sentence Prediction (NSP) task are optimal for relation classification.

Yilun Zhu, Sameer Pradhan, Amir Zeldes

SOTA coreference resolution produces increasingly impressive scores on the OntoNotes benchmark. However lack of comparable data following the same scheme for more genres makes it difficult to evaluate generalizability to open domain data. This paper provides a dataset and comprehensive evaluation showing that the latest neural LM based end-to-end systems degrade very substantially out of domain. We make an OntoNotes-like coreference dataset called OntoGUM publicly available, converted from GUM, an English corpus covering 12 genres, using deterministic rules, which we evaluate. Thanks to the rich syntactic and discourse annotations in GUM, we are able to create the largest human-annotated coreference corpus following the OntoNotes guidelines, and the first to be evaluated for consistency with the OntoNotes scheme. Out-of-domain evaluation across 12 genres shows nearly 15-20% degradation for both deterministic and deep learning systems, indicating a lack of generalizability or covert overfitting in existing coreference resolution models.

Luke Gessler, Siyao Peng, Yang Liu et al.

We present a freely available, genre-balanced English web corpus totaling 4M tokens and featuring a large number of high-quality automatic annotation layers, including dependency trees, non-named entity annotations, coreference resolution, and discourse trees in Rhetorical Structure Theory. By tapping open online data sources the corpus is meant to offer a more sizable alternative to smaller manually created annotated data sets, while avoiding pitfalls such as imbalanced or unknown composition, licensing problems, and low-quality natural language processing. We harness knowledge from multiple annotation layers in order to achieve a "better than NLP" benchmark and evaluate the accuracy of the resulting resource.

Siyao Peng, Yang Liu, Yilun Zhu et al.

Adpositions are frequent markers of semantic relations, but they are highly ambiguous and vary significantly from language to language. Moreover, there is a dearth of annotated corpora for investigating the cross-linguistic variation of adposition semantics, or for building multilingual disambiguation systems. This paper presents a corpus in which all adpositions have been semantically annotated in Mandarin Chinese; to the best of our knowledge, this is the first Chinese corpus to be broadly annotated with adposition semantics. Our approach adapts a framework that defined a general set of supersenses according to ostensibly language-independent semantic criteria, though its development focused primarily on English prepositions (Schneider et al., 2018). We find that the supersense categories are well-suited to Chinese adpositions despite syntactic differences from English. On a Mandarin translation of The Little Prince, we achieve high inter-annotator agreement and analyze semantic correspondences of adposition tokens in bitext.

Yue Yu, Yilun Zhu, Yang Liu et al.

In this paper we present GumDrop, Georgetown University's entry at the DISRPT 2019 Shared Task on automatic discourse unit segmentation and connective detection. Our approach relies on model stacking, creating a heterogeneous ensemble of classifiers, which feed into a metalearner for each final task. The system encompasses three trainable component stacks: one for sentence splitting, one for discourse unit segmentation and one for connective detection. The flexibility of each ensemble allows the system to generalize well to datasets of different sizes and with varying levels of homogeneity.

Yilun Zhu, Yang Liu, Siyao Peng et al.

This study adapts Semantic Network of Adposition and Case Supersenses (SNACS) annotation to Mandarin Chinese and demonstrates that the same supersense categories are appropriate for Chinese adposition semantics. We annotated 15 chapters of The Little Prince, with high interannotator agreement. The parallel corpus gives insight into differences in construal between the two languages' adpositions, namely a number of construals that are frequent in Chinese but rare or unattested in the English corpus. The annotated corpus can further support automatic disambiguation of adpositions in Chinese, and the common inventory of supersenses between the two languages can potentially serve cross-linguistic tasks such as machine translation.